ABSTRACT
National Center for Supercomputing Applications, University of Illinois at Urbana-Champaign
Jairo Panetta
Instituto Tecnologico de Aeronautica – ITA
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 150 8.2 Load Balancing Strategies for Weather Models . . . . . . . . . . . . . . . . . 151 8.3 The BRAMS Weather Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153 8.4 Load Balancing Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
8.4.1 Adaptations to AMPI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 155 8.4.2 Balancing Algorithms Employed . . . . . . . . . . . . . . . . . . . . . . . . 157
8.5 New Load Balancer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158 8.6 Fully Distributed Strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 162
8.6.1 Hilbert Curve-Based Load Balancer . . . . . . . . . . . . . . . . . . . . . 162 8.6.2 Diffusion-Based Load Balancer . . . . . . . . . . . . . . . . . . . . . . . . . . 164
8.7 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 166 8.7.1 First Set of Experiments: Privatization Strategy . . . . . . . 166 8.7.2 Second Set of Experiments: Virtualization Effects . . . . . . 167 8.7.3 Third Set of Experiments: Centralized Load Balancers . 171 8.7.4 Fourth Set of Experiments: Distributed Load Balancers 174
8.7.4.1 Centralized versus Distributed Load Balancers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 175
8.7.4.2 DiffusionLB versus HilbertLB . . . . . . . . . . . . . 178 8.8 Final Remarks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
Parallel Approach
Numerical weather forecasting models have a long history. Currently they have become important decision tools in many different areas, ranging from agriculture to flight control and energy production. Many applications require forecasts with very high resolution. To execute these models in a feasible amount of time, they are typically run in parallel.
